Time-Segmented Fraud Risk Scoring With Deep Learning and Survival Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing fraud detection systems in transaction service providers struggle to dynamically assess the changing likelihood of fraud over time, providing only a single probability value within a fixed time period and failing to indicate future fraud risks or account for varying fraud likelihoods at different times.
Innovation Solution
A method using deep learning models, such as RNNs or LSTM networks, combined with survival models like Cox regression, generates fraud risk scores for each subperiod following an attempted attack, providing a probability that a fraudulent transaction will not occur, allowing for dynamic fraud assessment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a deep learning model is used to predict fraud likelihood within a fixed time period, then a single probability value is provided, but the model does not indicate fraud likelihood at different times in the future or account for changing fraud likelihood at different subperiods
Solution Approach 1:
The patent divides the fixed time period into multiple subperiods and generates separate fraud risk scores for each subperiod. This segmentation allows the system to capture time-sensitive fraud patterns and provide granular predictions for different time intervals, resolving the contradiction by maintaining measurement precision while preserving time-sensitive information through structured division of the prediction horizon.
Solution Approach 2:
The patent implements dynamic fraud risk scoring by adjusting the fraud likelihood assessment for each subperiod based on changing patterns. The system dynamically adapts the prediction to reflect varying fraud risks at different times, transforming the static single-value prediction into a dynamic time-series of risk assessments that capture evolving fraud threats.
2Reliability
If a deep learning model is run periodically to assess fraud, then predictions are made, but the model is inefficient due to the amount of time and resources required to train and make predictions
Solution Approach 1:
The patent segments the model execution into incremental updates for each subperiod rather than requiring full periodic retraining. This allows the system to maintain reliable fraud predictions while significantly improving execution efficiency by processing smaller, time-specific data segments and updating predictions incrementally across subperiods.
Solution Approach 2:
The patent performs preliminary data processing and feature extraction for each subperiod in advance, preparing data segments that can be quickly processed by the deep learning model. This preliminary action reduces the computational burden during actual prediction execution, improving productivity while maintaining prediction reliability through pre-processed, ready-to-analyze data.
3Ease of operation
If static rules are used to assess fraud likelihood, then individual transaction fraud probability is determined, but the rules cannot be updated over time to account for different fraudster behavior
Solution Approach 1:
The patent replaces static rules with a dynamic deep learning model that automatically adapts to changing fraud patterns. The model learns from historical data across multiple subperiods and adjusts its predictions to reflect evolving fraudster behaviors, maintaining ease of operation through automated learning while achieving high adaptability to new fraud patterns.
Solution Approach 2:
The patent implements a feedback mechanism where the deep learning model continuously learns from actual fraud outcomes and updates its predictions for subsequent subperiods. This feedback loop enables the system to maintain simple operation through automated adjustments while becoming increasingly adaptable to new fraud patterns based on real-world performance data.
Data Source
AI summary
Provided is a method for fraud prevention using deep learning and survival models. The method may include receiving, with at least one processor, transaction data associated with a plurality of transactions of at least one payment account. At least one attempted attack may be detected based on the transaction data. A fraud risk score for each subperiod of a plurality of subperiods in a time period following the at least one attempted attack may be generated based on the transaction data using a deep learning model and a survival model. The fraud risk score for each respective subperiod may be associated with a probability that a fraudulent transaction will not occur by the respective subperiod. A system and computer program product are also disclosed.